Simple and Scalable Predictive UncertaintyEstimation using Deep Ensembles

Balaji Lakshminarayanan Alexander Pritzel Charles BlundellDeepMind

{balajiln,apritzel,cblundell}@google.com

Abstract

Deep neural networks (NNs) are powerful black box predictors that have recentlyachieved impressive performance on a wide spectrum of tasks. Quantifying pre-dictive uncertainty in NNs is a challenging and yet unsolved problem. BayesianNNs, which learn a distribution over weights, are currently the state-of-the-artfor estimating predictive uncertainty; however these require significant modifica-tions to the training procedure and are computationally expensive compared tostandard (non-Bayesian) NNs. We propose an alternative to Bayesian NNs thatis simple to implement, readily parallelizable, requires very little hyperparametertuning, and yields high quality predictive uncertainty estimates. Through a seriesof experiments on classification and regression benchmarks, we demonstrate thatour method produces well-calibrated uncertainty estimates which are as good orbetter than approximate Bayesian NNs. To assess robustness to dataset shift, weevaluate the predictive uncertainty on test examples from known and unknowndistributions, and show that our method is able to express higher uncertainty onout-of-distribution examples. We demonstrate the scalability of our method byevaluating predictive uncertainty estimates on ImageNet.

1 Introduction

Deep neural networks (NNs) have achieved state-of-the-art performance on a wide variety of machinelearning tasks [35] and are becoming increasingly popular in domains such as computer vision[32], speech recognition [25], natural language processing [42], and bioinformatics [2, 61]. Despiteimpressive accuracies in supervised learning benchmarks, NNs are poor at quantifying predictiveuncertainty, and tend to produce overconfident predictions. Overconfident incorrect predictions can beharmful or offensive [3], hence proper uncertainty quantification is crucial for practical applications.

Evaluating the quality of predictive uncertainties is challenging as the ‘ground truth’ uncertaintyestimates are usually not available. In this work, we shall focus upon two evaluation measures thatare motivated by practical applications of NNs. Firstly, we shall examine calibration [12, 13], afrequentist notion of uncertainty which measures the discrepancy between subjective forecasts and(empirical) long-run frequencies. The quality of calibration can be measured by proper scoring rules[17] such as log predictive probabilities and the Brier score [9]. Note that calibration is an orthogonalconcern to accuracy: a network’s predictions may be accurate and yet miscalibrated, and vice versa.The second notion of quality of predictive uncertainty we consider concerns generalization of thepredictive uncertainty to domain shift (also referred to as out-of-distribution examples [23]), that is,measuring if the network knows what it knows. For example, if a network trained on one dataset isevaluated on a completely different dataset, then the network should output high predictive uncertaintyas inputs from a different dataset would be far away from the training data. Well-calibrated predictionsthat are robust to model misspecification and dataset shift have a number of important practical uses(e.g., weather forecasting, medical diagnosis).

31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.

arX

iv:1

612.

0147

4v3

[st

at.M

L]

4 N

ov 2

017

There has been a lot of recent interest in adapting NNs to encompass uncertainty and probabilisticmethods. The majority of this work revolves around a Bayesian formalism [4], whereby a priordistribution is specified upon the parameters of a NN and then, given the training data, the posteriordistribution over the parameters is computed, which is used to quantify predictive uncertainty.Since exact Bayesian inference is computationally intractable for NNs, a variety of approximationshave been developed including Laplace approximation [40], Markov chain Monte Carlo (MCMC)methods [46], as well as recent work on variational Bayesian methods [6, 19, 39], assumed densityfiltering [24], expectation propagation [21, 38] and stochastic gradient MCMC variants such asLangevin diffusion methods [30, 59] and Hamiltonian methods [53]. The quality of predictiveuncertainty obtained using Bayesian NNs crucially depends on (i) the degree of approximation dueto computational constraints and (ii) if the prior distribution is ‘correct’, as priors of conveniencecan lead to unreasonable predictive uncertainties [50]. In practice, Bayesian NNs are often harderto implement and computationally slower to train compared to non-Bayesian NNs, which raisesthe need for a ‘general purpose solution’ that can deliver high-quality uncertainty estimates and yetrequires only minor modifications to the standard training pipeline.

Recently, Gal and Ghahramani [15] proposed using Monte Carlo dropout (MC-dropout) to estimatepredictive uncertainty by using Dropout [54] at test time. There has been work on approximateBayesian interpretation [15, 29, 41] of dropout. MC-dropout is relatively simple to implementleading to its popularity in practice. Interestingly, dropout may also be interpreted as ensemble modelcombination [54] where the predictions are averaged over an ensemble of NNs (with parametersharing). The ensemble interpretation seems more plausible particularly in the scenario where thedropout rates are not tuned based on the training data, since any sensible approximation to the trueBayesian posterior distribution has to depend on the training data. This interpretation motivates theinvestigation of ensembles as an alternative solution for estimating predictive uncertainty.

It has long been observed that ensembles of models improve predictive performance (see [14] for areview). However it is not obvious when and why an ensemble of NNs can be expected to producegood uncertainty estimates. Bayesian model averaging (BMA) assumes that the true model lies withinthe hypothesis class of the prior, and performs soft model selection to find the single best model withinthe hypothesis class [43]. In contrast, ensembles perform model combination, i.e. they combine themodels to obtain a more powerful model; ensembles can be expected to be better when the true modeldoes not lie within the hypothesis class. We refer to [11, 43] and [34, §2.5] for related discussions.It is important to note that even exact BMA is not guaranteed be robust to mis-specification withrespect to domain shift.

Summary of contributions: Our contribution in this paper is two fold. First, we describe a simple andscalable method for estimating predictive uncertainty estimates from NNs. We argue for trainingprobabilistic NNs (that model predictive distributions) using a proper scoring rule as the trainingcriteria. We additionally investigate the effect of two modifications to the training pipeline, namely(i) ensembles and (ii) adversarial training [18] and describe how they can produce smooth predictiveestimates. Secondly, we propose a series of tasks for evaluating the quality of the predictive uncertaintyestimates, in terms of calibration and generalization to unknown classes in supervised learningproblems. We show that our method significantly outperforms (or matches) MC-dropout. These tasks,along with our simple yet strong baseline, serve as an useful benchmark for comparing predictiveuncertainty estimates obtained using different Bayesian/non-Bayesian/hybrid methods.

Novelty and Significance: Ensembles of NNs, or deep ensembles for short, have been successfullyused to boost predictive performance (e.g. classification accuracy in ImageNet or Kaggle contests)and adversarial training has been used to improve robustness to adversarial examples. However, tothe best of our knowledge, ours is the first work to investigate their usefulness for predictive uncer-tainty estimation and compare their performance to current state-of-the-art approximate Bayesianmethods on a series of classification and regression benchmark datasets. Compared to BayesianNNs (e.g. variational inference or MCMC methods), our method is much simpler to implement,requires surprisingly few modifications to standard NNs, and well suited for distributed computation,thereby making it attractive for large-scale deep learning applications. To demonstrate scalability ofour method, we evaluate predictive uncertainty on ImageNet (and are the first to do so, to the best ofour knowledge). Most work on uncertainty in deep learning focuses on Bayesian deep learning; wehope that the simplicity and strong empirical performance of our approach will spark more interest innon-Bayesian approaches for predictive uncertainty estimation.

2

2 Deep Ensembles: A Simple Recipe For Predictive Uncertainty Estimation

2.1 Problem setup and High-level summary

We assume that the training dataset D consists of N i.i.d. data points D = {xn, yn}Nn=1, wherex ∈ RD represents the D-dimensional features. For classification problems, the label is assumedto be one of K classes, that is y ∈ {1, . . . ,K}. For regression problems, the label is assumed tobe real-valued, that is y ∈ R. Given the input features x, we use a neural network to model theprobabilistic predictive distribution pθ(y|x) over the labels, where θ are the parameters of the NN.We suggest a simple recipe: (1) use a proper scoring rule as the training criterion, (2) use adversarialtraining [18] to smooth the predictive distributions, and (3) train an ensemble. Let M denote thenumber of NNs in the ensemble and {θm}Mm=1 denote the parameters of the ensemble. We firstdescribe how to train a single neural net and then explain how to train an ensemble of NNs.

2.2 Proper scoring rules

Scoring rules measure the quality of predictive uncertainty (see [17] for a review). A scoring ruleassigns a numerical score to a predictive distribution pθ(y|x), rewarding better calibrated predictionsover worse. We shall consider scoring rules where a higher numerical score is better. Let a scoringrule be a function S(pθ, (y,x)) that evaluates the quality of the predictive distribution pθ(y|x) relativeto an event y|x ∼ q(y|x) where q(y,x) denotes the true distribution on (y,x)-tuples. The expectedscoring rule is then S(pθ, q) =

∫q(y,x)S(pθ, (y,x))dydx. A proper scoring rule is one where

S(pθ, q) ≤ S(q, q) with equality if and only if pθ(y|x) = q(y|x), for all pθ and q. NNs can then betrained according to measure that encourages calibration of predictive uncertainty by minimizing theloss L(θ) = −S(pθ, q).It turns out many common NN loss functions are proper scoring rules. For example, when maximizinglikelihood, the score function is S(pθ, (y,x)) = log pθ(y|x), and this is a proper scoring rule dueto Gibbs inequality: S(pθ, q) = Eq(x)q(y|x) log pθ(y|x) ≤ Eq(x)q(y|x) log q(y|x). In the case ofmulti-class K-way classification, the popular softmax cross entropy loss is equivalent to the loglikelihood and is a proper scoring rule. Interestingly, L(θ) = −S(pθ, (y,x)) = K−1

∑Kk=1

(δk=y −

pθ(y = k|x))2

, i.e., minimizing the squared error between the predictive probability of a label andone-hot encoding of the correct label, is also a proper scoring rule known as the Brier score [9].This provides justification for this common trick for training NNs by minimizing the squared errorbetween a binary label and its associated probability and shows it is, in fact, a well defined loss withdesirable properties.1

2.2.1 Training criterion for regression

For regression problems, NNs usually output a single value say µ(x) and the parameters are optimizedto minimize the mean squared error (MSE) on the training set, given by

∑Nn=1

(yn − µ(xn)

)2.

However, the MSE does not capture predictive uncertainty. Following [47], we use a networkthat outputs two values in the final layer, corresponding to the predicted mean µ(x) and variance2σ2(x) > 0. By treating the observed value as a sample from a (heteroscedastic) Gaussian distributionwith the predicted mean and variance, we minimize the negative log-likelihood criterion:

− log pθ(yn|xn) =log σ2θ(x)

2+

(y − µθ(x)

)22σ2θ(x)

+ constant. (1)

We found the above to perform satisfactorily in our experiments. However, two simple extensions areworth further investigation: (i) Maximum likelihood estimation over µθ(x) and σ2θ(x) might overfit;one could impose a prior and perform maximum-a-posteriori (MAP) estimation. (ii) In cases wherethe Gaussian is too-restrictive, one could use a complex distribution e.g. mixture density network [5]or a heavy-tailed distribution.

1Indeed as noted in Gneiting and Raftery [17], it can be shown that asymptotically maximizing any properscoring rule recovers true parameter values.

2We enforce the positivity constraint on the variance by passing the second output through the softplusfunction log(1 + exp(·)), and add a minimum variance (e.g. 10−6) for numerical stability.

3

2.3 Adversarial training to smooth predictive distributions

Adversarial examples, proposed by Szegedy et al. [55] and extended by Goodfellow et al. [18], arethose which are ‘close’ to the original training examples (e.g. an image that is visually indistin-guishable from the original image to humans), but are misclassified by the NN. Goodfellow et al.[18] proposed the fast gradient sign method as a fast solution to generate adversarial examples.Given an input x with target y, and loss `(θ,x, y) (e.g. − log pθ(y|x)), the fast gradient sign methodgenerates an adversarial example as x′ = x + � sign

(∇x `(θ,x, y)

), where � is a small value such

that the max-norm of the perturbation is bounded. Intuitively, the adversarial perturbation createsa new training example by adding a perturbation along a direction which the network is likely toincrease the loss. Assuming � is small enough, these adversarial examples can be used to augmentthe original training set by treating (x′, y) as additional training examples. This procedure, referredto as adversarial training,3 was found to improve the classifier’s robustness [18].

Interestingly, adversarial training can be interpreted as a computationally efficient solution to smooththe predictive distributions by increasing the likelihood of the target around an �-neighborhood ofthe observed training examples. Ideally one would want to smooth the predictive distributions alongall 2D directions in {1,−1}D; however this is computationally expensive. A random directionmight not necessarily increase the loss; however, adversarial training by definition computes thedirection where the loss is high and hence is better than a random direction for smoothing predictivedistributions. Miyato et al. [44] proposed a related idea called virtual adversarial training (VAT),where they picked ∆x = arg max∆x KL

(p(y|x)||p(y|x + ∆x)

); the advantage of VAT is that

it does not require knowledge of the true target y and hence can be applied to semi-supervisedlearning. Miyato et al. [44] showed that distributional smoothing using VAT is beneficial for efficientsemi-supervised learning; in contrast, we investigate the use of adversarial training for predictiveuncertainty estimation. Hence, our contributions are complementary; one could use VAT or otherforms of adversarial training, cf. [33], for improving predictive uncertainty in the semi-supervisedsetting as well.

2.4 Ensembles: training and prediction

The most popular ensembles use decision trees as the base learners and a wide variety of methodhave been explored in the literature on ensembles. Broadly, there are two classes of ensembles:randomization-based approaches such as random forests [8], where the ensemble members canbe trained in parallel without any interaction, and boosting-based approaches where the ensemblemembers are fit sequentially. We focus only on the randomization based approach as it is better suitedfor distributed, parallel computation. Breiman [8] showed that the generalization error of randomforests can be upper bounded by a function of the strength and correlation between individual trees;hence it is desirable to use a randomization scheme that de-correlates the predictions of the individualmodels as well as ensures that the individual models are strong (e.g. high accuracy). One of thepopular strategies is bagging (a.k.a. bootstrapping), where ensemble members are trained on differentbootstrap samples of the original training set. If the base learner lacks intrinsic randomization (e.g. itcan be trained efficiently by solving a convex optimization problem), bagging is a good mechanismfor inducing diversity. However, if the underlying base learner has multiple local optima, as is thecase typically with NNs, the bootstrap can sometimes hurt performance since a base learner trainedon a bootstrap sample sees only 63% unique data points.4 In the literature on decision tree ensembles,Breiman [8] proposed to use a combination of bagging [7] and random subset selection of features ateach node. Geurts et al. [16] later showed that bagging is unnecessary if additional randomness canbe injected into the random subset selection procedure. Intuitively, using more data for training thebase learners helps reduce their bias and ensembling helps reduce the variance.

We used the entire training dataset to train each network since deep NNs typically perform betterwith more data, although it is straightforward to use a random subsample if need be. We found thatrandom initialization of the NN parameters, along with random shuffling of the data points, wassufficient to obtain good performance in practice. We observed that bagging deteriorated performancein our experiments. Lee et al. [36] independently observed that training on entire dataset withrandom initialization was better than bagging for deep ensembles, however their goal was to improve

3Not to be confused with Generative Adversarial Networks (GANs).4 The bootstrap draws N times uniformly with replacement from a dataset with N items. The probability

an item is picked at least once is 1− (1− 1/N)N , which for large N becomes 1− e−1 ≈ 0.632. Hence, thenumber of unique data points in a bootstrap sample is 0.632×N on average.

4

predictive accuracy and not predictive uncertainty. The overall training procedure is summarized inAlgorithm 1.

Algorithm 1 Pseudocode of the training procedure for our method1: . Let each neural network parametrize a distribution over the outputs, i.e. pθ(y|x). Use a proper

scoring rule as the training criterion `(θ,x, y). Recommended default values are M = 5 and� = 1% of the input range of the corresponding dimension (e.g 2.55 if input range is [0,255]).

2: Initialize θ1, θ2, . . . , θM randomly3: for m = 1 : M do . train networks independently in parallel4: Sample data point nm randomly for each net . single nm for clarity, minibatch in practice5: Generate adversarial example using x′nm = xnm + � sign

(∇xnm `(θm,xnm , ynm)

)6: Minimize `(θm,xnm , ynm) + `(θm,x

′nm , ynm) w.r.t. θm . adversarial training (optional)

We treat the ensemble as a uniformly-weighted mixture model and combine the predictions asp(y|x) = M−1∑Mm=1 pθm(y|x, θm). For classification, this corresponds to averaging the predictedprobabilities. For regression, the prediction is a mixture of Gaussian distributions. For ease ofcomputing quantiles and predictive probabilities, we further approximate the ensemble prediction as aGaussian whose mean and variance are respectively the mean and variance of the mixture. The meanand variance of a mixture M−1

∑N (µθm(x), σ2θm(x)) are given by µ∗(x) = M−1∑m µθm(x)and σ2∗(x) = M

−1∑m

(σ2θm(x) + µ

2θm

(x))− µ2∗(x) respectively.

3 Experimental results

3.1 Evaluation metrics and experimental setup

For both classification and regression, we evaluate the negative log likelihood (NLL) which dependson the predictive uncertainty. NLL is a proper scoring rule and a popular metric for evaluatingpredictive uncertainty [49]. For classification we additionally measure classification accuracy andthe Brier score, defined as BS = K−1

∑Kk=1

(t∗k − p(y = k|x∗)

)2where t∗k = 1 if k = y

∗, and 0otherwise. For regression problems, we additionally measured the root mean squared error (RMSE).Unless otherwise specified, we used batch size of 100 and Adam optimizer with fixed learning rate of0.1 in our experiments. We use the same technique for generating adversarial training examples forregression problems. Goodfellow et al. [18] used a fixed � for all dimensions; this is unsatisfyingif the input dimensions have different ranges. Hence, in all of our experiments, we set � to 0.01times the range of the training data along that particular dimension. We used the default weightinitialization in Torch.

3.2 Regression on toy datasets

First, we qualitatively evaluate the performance of the proposed method on a one-dimensional toyregression dataset. This dataset was used by Hernández-Lobato and Adams [24], and consists of 20training examples drawn as y = x3 + � where � ∼ N (0, 32). We used the same architecture as [24].A commonly used heuristic in practice is to use an ensemble of NNs (trained to minimize MSE),obtain multiple point predictions and use the empirical variance of the networks’ predictions as anapproximate measure of uncertainty. We demonstrate that this is inferior to learning the variance bytraining using NLL.5 The results are shown in Figure 1.

The results clearly demonstrate that (i) learning variance and training using a scoring rule (NLL) leadsto improved predictive uncertainty and (ii) ensemble combination improves performance, especiallyas we move farther from the observed training data.

3.3 Regression on real world datasets

In our next experiment, we compare our method to state-of-the-art methods for predictive uncertaintyestimation using NNs on regression tasks. We use the experimental setup proposed by Hernández-Lobato and Adams [24] for evaluating probabilistic backpropagation (PBP), which was also used

5See also Appendix A.2 for calibration results on a real world dataset.

5

−6 −4 −2 0 2 4 6−200

−100

0

100

200

−6 −4 −2 0 2 4 6−200

−100

0

100

200

−6 −4 −2 0 2 4 6−200

−100

0

100

200

−6 −4 −2 0 2 4 6−300−200−100

0

100

200

300

Figure 1: Results on a toy regression task: x-axis denotes x. On the y-axis, the blue line is the groundtruth curve, the red dots are observed noisy training data points and the gray lines correspond tothe predicted mean along with three standard deviations. Left most plot corresponds to empiricalvariance of 5 networks trained using MSE, second plot shows the effect of training using NLL usinga single net, third plot shows the additional effect of adversarial training, and final plot shows theeffect of using an ensemble of 5 networks respectively.

by Gal and Ghahramani [15] to evaluate MC-dropout.6 Each dataset is split into 20 train-test folds,except for the protein dataset which uses 5 folds and the Year Prediction MSD dataset which usesa single train-test split. We use the identical network architecture: 1-hidden layer NN with ReLUnonlinearity [45], containing 50 hidden units for smaller datasets and 100 hidden units for the largerprotein and Year Prediction MSD datasets. We trained for 40 epochs; we refer to [24] for furtherdetails about the datasets and the experimental protocol. We used 5 networks in our ensemble. Ourresults are shown in Table 1, along with the PBP and MC-dropout results reported in their respectivepapers.

Datasets RMSE NLLPBP MC-dropout Deep Ensembles PBP MC-dropout Deep Ensembles

Boston housing 3.01 ± 0.18 2.97 ± 0.85 3.28 ± 1.00 2.57 ± 0.09 2.46 ± 0.25 2.41 ± 0.25Concrete 5.67 ± 0.09 5.23 ± 0.53 6.03 ± 0.58 3.16 ± 0.02 3.04 ± 0.09 3.06 ± 0.18Energy 1.80 ± 0.05 1.66 ± 0.19 2.09 ± 0.29 2.04 ± 0.02 1.99 ± 0.09 1.38 ± 0.22Kin8nm 0.10 ± 0.00 0.10 ± 0.00 0.09 ± 0.00 -0.90 ± 0.01 -0.95 ± 0.03 -1.20 ± 0.02Naval propulsion plant 0.01 ± 0.00 0.01 ± 0.00 0.00 ± 0.00 -3.73 ± 0.01 -3.80 ± 0.05 -5.63 ± 0.05Power plant 4.12 ± 0.03 4.02 ± 0.18 4.11 ± 0.17 2.84 ± 0.01 2.80 ± 0.05 2.79 ± 0.04Protein 4.73 ± 0.01 4.36 ± 0.04 4.71 ± 0.06 2.97 ± 0.00 2.89 ± 0.01 2.83 ± 0.02Wine 0.64 ± 0.01 0.62 ± 0.04 0.64 ± 0.04 0.97 ± 0.01 0.93 ± 0.06 0.94 ± 0.12Yacht 1.02 ± 0.05 1.11 ± 0.38 1.58 ± 0.48 1.63 ± 0.02 1.55 ± 0.12 1.18 ± 0.21Year Prediction MSD 8.88 ± NA 8.85 ± NA 8.89 ± NA 3.60 ± NA 3.59 ± NA 3.35 ± NA

Table 1: Results on regression benchmark datasets comparing RMSE and NLL. See Table 2 forresults on variants of our method.

We observe that our method outperforms (or is competitive with) existing methods in terms of NLL.On some datasets, we observe that our method is slightly worse in terms of RMSE. We believe thatthis is due to the fact that our method optimizes for NLL (which captures predictive uncertainty)instead of MSE. Table 2 in Appendix A.1 reports additional results on variants of our method,demonstrating the advantage of using an ensemble as well as learning variance.

3.4 Classification on MNIST, SVHN and ImageNet

Next we evaluate the performance on classification tasks using MNIST and SVHN datasets. Our goalis not to achieve the state-of-the-art performance on these problems, but rather to evaluate the effectof adversarial training as well as the number of networks in the ensemble. To verify if adversarialtraining helps, we also include a baseline which picks a random signed vector. For MNIST, we usedan MLP with 3-hidden layers with 200 hidden units per layer and ReLU non-linearities with batchnormalization. For MC-dropout, we added dropout after each non-linearity with 0.1 as the dropoutrate.7 Results are shown in Figure 2(a). We observe that adversarial training and increasing thenumber of networks in the ensemble significantly improve performance in terms of both classificationaccuracy as well as NLL and Brier score, illustrating that our method produces well-calibrateduncertainty estimates. Adversarial training leads to better performance than augmenting with randomdirection. Our method also performs much better than MC-dropout in terms of all the performancemeasures. Note that augmenting the training dataset with invariances (such as random crop andhorizontal flips) is complementary to adversarial training and can potentially improve performance.

6We do not compare to VI [19] as PBP and MC-dropout outperform VI on these benchmarks.7We also tried dropout rate of 0.5, but that performed worse.

6

0 5 10 15Number of nets

1.0

1.2

1.4

1.6

1.8Classification Error

Ensemble

Ensemble + R

Ensemble + AT

MC dropout

0 5 10 15Number of nets

0.02

0.04

0.06

0.08

0.10

0.12

0.14NLL

Ensemble

Ensemble + R

Ensemble + AT

MC dropout

0 5 10 15Number of nets

0.0014

0.0016

0.0018

0.0020

0.0022

0.0024

0.0026

0.0028

0.0030Brier Score

Ensemble

Ensemble + R

Ensemble + AT

MC dropout

(a) MNIST dataset using 3-layer MLP

0 5 10Number of nets

2

4

6

8

10

12

14Classification Error

Ensemble

Ensemble + R

Ensemble + AT

MC dropout

0 5 10Number of nets

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50NLL

Ensemble

Ensemble + R

Ensemble + AT

MC dropout

0 5 10Number of nets

0.004

0.006

0.008

0.010

0.012

0.014

0.016Brier Score

Ensemble

Ensemble + R

Ensemble + AT

MC dropout

(b) SVHN using VGG-style convnet

Figure 2: Evaluating predictive uncertainty as a function of ensemble size M (number of networksin the ensemble or the number of MC-dropout samples): Ensemble variants significantly outperformMC-dropout performance with the corresponding M in terms of all 3 metrics. Adversarial trainingimproves results for MNIST for all M and SVHN when M = 1, but the effect drops as M increases.

To measure the sensitivity of the results to the choice of network architecture, we experimentedwith a two-layer MLP as well as a convolutional NN; we observed qualitatively similar results; seeAppendix B.1 in the supplementary material for details.

We also report results on the SVHN dataset using an VGG-style convolutional NN.8 The results arein Figure 2(b). Ensembles outperform MC dropout. Adversarial training helps slightly for M = 1,however the effect drops as the number of networks in the ensemble increases. If the classes arewell-separated, adversarial training might not change the classification boundary significantly. It isnot clear if this is the case here, further investigation is required.

Finally, we evaluate on the ImageNet (ILSVRC-2012) dataset [51] using the inception network [56].Due to computational constraints, we only evaluate the effect of ensembles on this dataset. Theresults on ImageNet (single-crop evaluation) are shown in Table 4. We observe that as M increases,both the accuracy and the quality of predictive uncertainty improve significantly.

Another advantage of using an ensemble is that it enables us to easily identify training exampleswhere the individual networks disagree or agree the most. This disagreement9 provides anotheruseful qualitative way to evaluate predictive uncertainty. Figures 10 and 11 in Appendix B.2 reportqualitative evaluation of predictive uncertainty on the MNIST dataset.

3.5 Uncertainty evaluation: test examples from known vs unknown classes

In the final experiment, we evaluate uncertainty on out-of-distribution examples from unseen classes.Overconfident predictions on unseen classes pose a challenge for reliable deployment of deep learningmodels in real world applications. We would like the predictions to exhibit higher uncertainty whenthe test data is very different from the training data. To test if the proposed method possesses thisdesirable property, we train a MLP on the standard MNIST train/test split using the same architectureas before. However, in addition to the regular test set with known classes, we also evaluate it on atest set containing unknown classes. We used the test split of the NotMNIST10 dataset. The imagesin this dataset have the same size as MNIST, however the labels are alphabets instead of digits. Wedo not have access to the true conditional probabilities, but we expect the predictions to be closerto uniform on unseen classes compared to the known classes where the predictive probabilitiesshould concentrate on the true targets. We evaluate the entropy of the predictive distribution anduse this to evaluate the quality of the uncertainty estimates. The results are shown in Figure 3(a).For known classes (top row), both our method and MC-dropout have low entropy as expected. Forunknown classes (bottom row), as M increases, the entropy of deep ensembles increases much fasterthan MC-dropout indicating that our method is better suited for handling unseen test examples. Inparticular, MC-dropout seems to give high confidence predictions for some of the test examples, asevidenced by the mode around 0 even for unseen classes. Such overconfident wrong predictions canbe problematic in practice when tested on a mixture of known and unknown classes, as we will see inSection 3.6. Comparing different variants of our method, the mode for adversarial training increasesslightly faster than the mode for vanilla ensembles indicating that adversarial training is beneficial

8The architecture is similar to the one described in http://torch.ch/blog/2015/07/30/cifar.html.9More precisely, we define disagreement as

∑Mm=1 KL(pθm(y|x)||pE(y|x)) where KL denotes the

Kullback-Leibler divergence and pE(y|x) = M−1∑m pθm(y|x) is the prediction of the ensemble.

10Available at http://yaroslavvb.blogspot.co.uk/2011/09/notmnist-dataset.html

7

http://torch.ch/blog/2015/07/30/cifar.htmlhttp://yaroslavvb.blogspot.co.uk/2011/09/notmnist-dataset.html

for quantifying uncertainty on unseen classes. We qualitatively evaluate results in Figures 12(a)and 12(b) in Appendix B.2. Figure 12(a) shows that the ensemble agreement is highest for letter ‘I’which resembles 1 in the MNIST training dataset, and that the ensemble disagreement is higher forexamples visually different from the MNIST training dataset.

−0.5 0.0 0.5 1.0 1.5 2.0entropy values

0

2

4

6

8

10

12

14Ensemble

1

5

10

−0.5 0.0 0.5 1.0 1.5 2.0entropy values

Ensemble + R

1

5

10

−0.5 0.0 0.5 1.0 1.5 2.0entropy values

Ensemble + AT

1

5

10

−0.5 0.0 0.5 1.0 1.5 2.0entropy values

MC dropout 0.1

1

5

10

−0.5 0.0 0.5 1.0 1.5 2.0entropy values

0

2

4

6

8

10

12

14Ensemble

1

5

10

−0.5 0.0 0.5 1.0 1.5 2.0entropy values

Ensemble + R

1

5

10

−0.5 0.0 0.5 1.0 1.5 2.0entropy values

Ensemble + AT

1

5

10

−0.5 0.0 0.5 1.0 1.5 2.0entropy values

MC dropout 0.1

1

5

10

(a) MNIST-NotMNIST

−0.50.0 0.5 1.0 1.5 2.0 2.5entropy values

0

1

2

3

4

5

6

7Ensemble

1

5

10

−0.50.0 0.5 1.0 1.5 2.0 2.5entropy values

Ensemble + R

1

5

10

−0.50.0 0.5 1.0 1.5 2.0 2.5entropy values

Ensemble + AT

1

5

10

−0.5 0.0 0.5 1.0 1.5 2.0entropy values

MC dropout

1

5

10

−0.50.0 0.5 1.0 1.5 2.0 2.5entropy values

0

1

2

3

4

5

6

7Ensemble

1

5

10

−0.50.0 0.5 1.0 1.5 2.0 2.5entropy values

Ensemble + R

1

5

10

−0.50.0 0.5 1.0 1.5 2.0 2.5entropy values

Ensemble + AT

1

5

10

−0.50.0 0.5 1.0 1.5 2.0 2.5entropy values

MC dropout

1

5

10

(b) SVHN-CIFAR10

Figure 3: : Histogram of the predictive entropy on test examples from known classes (top row) andunknown classes (bottom row), as we vary ensemble size M .

We ran a similar experiment, training on SVHN and testing on CIFAR-10 [31] test set; both datasetscontain 32× 32× 3 images, however SVHN contains images of digits whereas CIFAR-10 containsimages of object categories. The results are shown in Figure 3(b). As in the MNIST-NotMNISTexperiment, we observe that MC-dropout produces over-confident predictions on unseen examples,whereas our method produces higher uncertainty on unseen classes.

Finally, we test on ImageNet by splitting the training set by categories. We split the dataset intoimages of dogs (known classes) and non-dogs (unknown classes), following Vinyals et al. [58] whoproposed this setup for a different task. Figure 5 shows the histogram of the predictive entropy aswell as the maximum predicted probability (i.e. confidence in the predicted class). We observe thatthe predictive uncertainty improves on unseen classes, as the ensemble size increases.

3.6 Accuracy as a function of confidence

In practical applications, it is highly desirable for a system to avoid overconfident, incorrect predictionsand fail gracefully. To evaluate the usefulness of predictive uncertainty for decision making, weconsider a task where the model is evaluated only on cases where the model’s confidence is above anuser-specified threshold. If the confidence estimates are well-calibrated, one can trust the model’spredictions when the reported confidence is high and resort to a different solution (e.g. use human ina loop, or use prediction from a simpler model) when the model is not confident.

We re-use the results from the experiment in the previous section where we trained a network onMNIST and test it on a mix of test examples from MNIST (known classes) and NotMNIST (unknown

M Top-1 error Top-5 error NLL Brier Score% % ×10−3

1 22.166 6.129 0.959 0.3172 20.462 5.274 0.867 0.2943 19.709 4.955 0.836 0.2864 19.334 4.723 0.818 0.2825 19.104 4.637 0.809 0.2806 18.986 4.532 0.803 0.2787 18.860 4.485 0.797 0.2778 18.771 4.430 0.794 0.2769 18.728 4.373 0.791 0.276

10 18.675 4.364 0.789 0.275

Figure 4: Results on ImageNet: DeepEnsembles lead to lower classificationerror as well as better predictive uncer-tainty as evidenced by lower NLL andBrier score.

0.20.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6

entropy values

0

1

2

3

4

5

6

7

8Known classes

1

2

3

4

5

1 0 1 2 3 4 5

entropy values

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7Unknown classes

1

2

3

4

5

0.2 0.4 0.6 0.8 1.0 1.2

max predicted prob

0

1

2

3

4

5

6

7

8Known classes

1

2

3

4

5

0.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2

max predicted prob

0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0Unknown classes

1

2

3

4

5

Figure 5: ImageNet trained only on dogs: Histogram of thepredictive entropy (left) and maximum predicted probabil-ity (right) on test examples from known classes (dogs) andunknown classes (non-dogs), as we vary the ensemble size.

8

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9Confidence Threshold ¿

30

40

50

60

70

80

90

Acc

ura

cy o

n e

xam

ple

s p(yjx)¸¿ Ensemble

Ensemble + R

Ensemble + AT

MC dropout

Figure 6: Accuracy vs Confidence curves: Networks trained on MNIST and tested on both MNISTtest containing known classes and the NotMNIST dataset containing unseen classes. MC-dropout canproduce overconfident wrong predictions, whereas deep ensembles are significantly more robust.

classes). The network will produce incorrect predictions on out-of-distribution examples, however wewould like these predictions to have low confidence. Given the prediction p(y = k|x), we define thepredicted label as ŷ = arg maxk p(y = k|x), and the confidence as p(y = ŷ|x) = maxk p(y = k|x).We filter out test examples, corresponding to a particular confidence threshold 0 ≤ τ ≤ 1 and plot theaccuracy for this threshold. The confidence vs accuracy results are shown in Figure 6. If we look atcases only where the confidence is ≥ 90%, we expect higher accuracy than cases where confidence≥ 80%, hence the curve should be monotonically increasing. If the application demands an accuracyx%, we can trust the model only in cases where the confidence is greater than the correspondingthreshold. Hence, we can compare accuracy of the models for a desired confidence threshold of theapplication. MC-dropout can produce overconfident wrong predictions as evidenced by low accuracyeven for high values of τ , whereas deep ensembles are significantly more robust.

4 Discussion

We have proposed a simple and scalable non-Bayesian solution that provides a very strong baselineon evaluation metrics for predictive uncertainty quantification. Intuitively, our method captures twosources of uncertainty. Training a probabilistic NN pθ(y|x) using proper scoring rules as trainingobjectives captures ambiguity in targets y for a given x. In addition, our method uses a combinationof ensembles (which captures “model uncertainty” by averaging predictions over multiple modelsconsistent with the training data), and adversarial training (which encourages local smoothness),for robustness to model misspecification and out-of-distribution examples. Ensembles, even forM = 5, significantly improve uncertainty quality in all the cases. Adversarial training helps onsome datasets for some metrics and is not strictly necessary in all cases. Our method requires verylittle hyperparameter tuning and is well suited for large scale distributed computation and can bereadily implemented for a wide variety of architectures such as MLPs, CNNs, etc including thosewhich do not use dropout e.g. residual networks [22]. It is perhaps surprising to the Bayesian deeplearning community that a non-Bayesian (yet probabilistic) approach can perform as well as BayesianNNs. We hope that our work will encourage the community to consider non-Bayesian approaches(such as ensembles) and other interesting evaluation metrics for predictive uncertainty. Concurrentwith our work, Hendrycks and Gimpel [23] and Guo et al. [20] have also independently shown thatnon-Bayesian solutions can produce good predictive uncertainty estimates on some tasks. Abbasiand Gagné [1], Tramèr et al. [57] have also explored ensemble-based solutions to tackle adversarialexamples, a particularly hard case of out-of-distribution examples.

There are several avenues for future work. We focused on training independent networks as trainingcan be trivially parallelized. Explicitly de-correlating networks’ predictions, e.g. as in [37], mightpromote ensemble diversity and improve performance even further. Optimizing the ensemble weights,as in stacking [60] or adaptive mixture of experts [28], can further improve the performance. Theensemble has M times more parameters than a single network; for memory-constrained applications,the ensemble can be distilled into a simpler model [10, 26]. It would be also interesting to investigateso-called implicit ensembles the where ensemble members share parameters, e.g. using multipleheads [36, 48], snapshot ensembles [27] or swapout [52].

9

Acknowledgments

We would like to thank Samuel Ritter and Oriol Vinyals for help with ImageNet experiments, andDaan Wierstra, David Silver, David Barrett, Ian Osband, Martin Szummer, Peter Dayan, ShakirMohamed, Theophane Weber, Ulrich Paquet and the anonymous reviewers for helpful feedback.

References

[1] M. Abbasi and C. Gagné. Robustness to adversarial examples through an ensemble of specialists.arXiv preprint arXiv:1702.06856, 2017.

[2] B. Alipanahi, A. Delong, M. T. Weirauch, and B. J. Frey. Predicting the sequence specificitiesof DNA-and RNA-binding proteins by deep learning. Nature biotechnology, 33(8):831–838,2015.

[3] D. Amodei, C. Olah, J. Steinhardt, P. Christiano, J. Schulman, and D. Mané. Concrete problemsin AI safety. arXiv preprint arXiv:1606.06565, 2016.

[4] J. M. Bernardo and A. F. Smith. Bayesian Theory, volume 405. John Wiley & Sons, 2009.

[5] C. M. Bishop. Mixture density networks. 1994.

[6] C. Blundell, J. Cornebise, K. Kavukcuoglu, and D. Wierstra. Weight uncertainty in neuralnetworks. In ICML, 2015.

[7] L. Breiman. Bagging predictors. Machine learning, 24(2):123–140, 1996.

[8] L. Breiman. Random forests. Machine learning, 45(1):5–32, 2001.

[9] G. W. Brier. Verification of forecasts expressed in terms of probability. Monthly weather review,1950.

[10] C. Bucila, R. Caruana, and A. Niculescu-Mizil. Model compression. In KDD. ACM, 2006.

[11] B. Clarke. Comparing Bayes model averaging and stacking when model approximation errorcannot be ignored. J. Mach. Learn. Res. (JMLR), 4:683–712, 2003.

[12] A. P. Dawid. The well-calibrated Bayesian. Journal of the American Statistical Association,1982.

[13] M. H. DeGroot and S. E. Fienberg. The comparison and evaluation of forecasters. Thestatistician, 1983.

[14] T. G. Dietterich. Ensemble methods in machine learning. In Multiple classifier systems. 2000.

[15] Y. Gal and Z. Ghahramani. Dropout as a Bayesian approximation: Representing modeluncertainty in deep learning. In ICML, 2016.

[16] P. Geurts, D. Ernst, and L. Wehenkel. Extremely randomized trees. Machine learning, 63(1):3–42, 2006.

[17] T. Gneiting and A. E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journalof the American Statistical Association, 102(477):359–378, 2007.

[18] I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. InICLR, 2015.

[19] A. Graves. Practical variational inference for neural networks. In NIPS, 2011.

[20] C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger. On calibration of modern neural networks.arXiv preprint arXiv:1706.04599, 2017.

[21] L. Hasenclever, S. Webb, T. Lienart, S. Vollmer, B. Lakshminarayanan, C. Blundell, and Y. W.Teh. Distributed Bayesian learning with stochastic natural-gradient expectation propagation andthe posterior server. arXiv preprint arXiv:1512.09327, 2015.

10

[22] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages770–778, 2016.

[23] D. Hendrycks and K. Gimpel. A baseline for detecting misclassified and out-of-distributionexamples in neural networks. arXiv preprint arXiv:1610.02136, 2016.

[24] J. M. Hernández-Lobato and R. P. Adams. Probabilistic backpropagation for scalable learningof Bayesian neural networks. In ICML, 2015.

[25] G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke,P. Nguyen, T. N. Sainath, et al. Deep neural networks for acoustic modeling in speech recog-nition: The shared views of four research groups. Signal Processing Magazine, IEEE, 29(6):82–97, 2012.

[26] G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprintarXiv:1503.02531, 2015.

[27] G. Huang, Y. Li, G. Pleiss, Z. Liu, J. E. Hopcroft, and K. Q. Weinberger. Snapshot ensembles:Train 1, get M for free. ICLR submission, 2017.

[28] R. A. Jacobs, M. I. Jordan, S. J. Nowlan, and G. E. Hinton. Adaptive mixtures of local experts.Neural computation, 3(1):79–87, 1991.

[29] D. P. Kingma, T. Salimans, and M. Welling. Variational dropout and the local reparameterizationtrick. In NIPS, 2015.

[30] A. Korattikara, V. Rathod, K. Murphy, and M. Welling. Bayesian dark knowledge. In NIPS,2015.

[31] A. Krizhevsky. Learning multiple layers of features from tiny images. 2009.

[32] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutionalneural networks. In NIPS, 2012.

[33] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial machine learning at scale. arXiv preprintarXiv:1611.01236, 2016.

[34] B. Lakshminarayanan. Decision trees and forests: a probabilistic perspective. PhD thesis, UCL(University College London), 2016.

[35] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 521(7553):436–444, 2015.

[36] S. Lee, S. Purushwalkam, M. Cogswell, D. Crandall, and D. Batra. Why M heads are better thanone: Training a diverse ensemble of deep networks. arXiv preprint arXiv:1511.06314, 2015.

[37] S. Lee, S. P. S. Prakash, M. Cogswell, V. Ranjan, D. Crandall, and D. Batra. Stochastic multiplechoice learning for training diverse deep ensembles. In NIPS, 2016.

[38] Y. Li, J. M. Hernández-Lobato, and R. E. Turner. Stochastic expectation propagation. In NIPS,2015.

[39] C. Louizos and M. Welling. Structured and efficient variational deep learning with matrixGaussian posteriors. arXiv preprint arXiv:1603.04733, 2016.

[40] D. J. MacKay. Bayesian methods for adaptive models. PhD thesis, California Institute ofTechnology, 1992.

[41] S.-i. Maeda. A Bayesian encourages dropout. arXiv preprint arXiv:1412.7003, 2014.

[42] T. Mikolov, K. Chen, G. Corrado, and J. Dean. Efficient estimation of word representations invector space. arXiv preprint arXiv:1301.3781, 2013.

[43] T. P. Minka. Bayesian model averaging is not model combination. 2000.

11

[44] T. Miyato, S.-i. Maeda, M. Koyama, K. Nakae, and S. Ishii. Distributional smoothing by virtualadversarial examples. In ICLR, 2016.

[45] V. Nair and G. E. Hinton. Rectified linear units improve restricted Boltzmann machines. InICML, 2010.

[46] R. M. Neal. Bayesian Learning for Neural Networks. Springer-Verlag New York, Inc., 1996.

[47] D. A. Nix and A. S. Weigend. Estimating the mean and variance of the target probabilitydistribution. In IEEE International Conference on Neural Networks, 1994.

[48] I. Osband, C. Blundell, A. Pritzel, and B. Van Roy. Deep exploration via bootstrapped DQN. InNIPS, 2016.

[49] J. Quinonero-Candela, C. E. Rasmussen, F. Sinz, O. Bousquet, and B. Schölkopf. Evaluatingpredictive uncertainty challenge. In Machine Learning Challenges. Springer, 2006.

[50] C. E. Rasmussen and J. Quinonero-Candela. Healing the relevance vector machine throughaugmentation. In ICML, 2005.

[51] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy,A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual RecognitionChallenge. International Journal of Computer Vision (IJCV), 115(3):211–252, 2015.

[52] S. Singh, D. Hoiem, and D. Forsyth. Swapout: Learning an ensemble of deep architectures. InNIPS, 2016.

[53] J. T. Springenberg, A. Klein, S. Falkner, and F. Hutter. Bayesian optimization with robustBayesian neural networks. In Advances in Neural Information Processing Systems, pages4134–4142, 2016.

[54] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: A simpleway to prevent neural networks from overfitting. JMLR, 2014.

[55] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus.Intriguing properties of neural networks. In ICLR, 2014.

[56] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethinking the inception archi-tecture for computer vision. In Proceedings of the IEEE Conference on Computer Vision andPattern Recognition, pages 2818–2826, 2016.

[57] F. Tramèr, A. Kurakin, N. Papernot, D. Boneh, and P. McDaniel. Ensemble adversarial training:Attacks and defenses. arXiv preprint arXiv:1705.07204, 2017.

[58] O. Vinyals, C. Blundell, T. Lillicrap, D. Wierstra, et al. Matching networks for one shot learning.In NIPS, 2016.

[59] M. Welling and Y. W. Teh. Bayesian learning via stochastic gradient Langevin dynamics. InICML, 2011.

[60] D. H. Wolpert. Stacked generalization. Neural networks, 5(2):241–259, 1992.

[61] J. Zhou and O. G. Troyanskaya. Predicting effects of noncoding variants with deep learning-based sequence model. Nature methods, 12(10):931–934, 2015.

12

Supplementary materialSimple and Scalable Predictive Uncertainty Estimation using Deep Ensembles

A Additional results on regression

A.1 Additional results on regression benchmarks

Datasets Ensemble-5 (MSE) Ensemble-10 (MSE) ML-1 ML-1 + AT ML-5

Boston housing 3.09 ± 0.84 3.10 ± 0.83 3.17 ± 1.00 3.18 ± 0.99 3.28 ± 1.00Concrete 5.73 ± 0.50 5.76 ± 0.50 6.08 ± 0.56 6.09 ± 0.54 6.03 ± 0.58Energy 1.61 ± 0.19 1.62 ± 0.18 2.11 ± 0.30 2.09 ± 0.29 2.09 ± 0.29Kin8nm 0.08 ± 0.00 0.08 ± 0.00 0.09 ± 0.00 0.09 ± 0.00 0.09 ± 0.00Naval propulsion plant 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00Power plant 4.08 ± 0.15 4.07 ± 0.15 4.10 ± 0.15 4.10 ± 0.15 4.11 ± 0.17Protein 4.49 ± 0.04 4.50 ± 0.02 4.64 ± 0.01 4.75 ± 0.11 4.71 ± 0.17Wine 0.64 ± 0.04 0.64 ± 0.04 0.64 ± 0.04 0.64 ± 0.04 0.64 ± 0.04Yacht 2.78 ± 0.59 2.68 ± 0.57 1.43 ± 0.57 1.47 ± 0.58 1.58 ± 0.48Year Prediction MSD 8.92 ± nan 8.95 ± nan 8.89 ± nan 9.02 ± nan 8.89 ± nanDatasets Ensemble-5 (MSE) Ensemble-10 (MSE) ML-1 ML-1 + AT ML-5

Boston housing 17.28 ± 6.17 10.61 ± 4.37 2.55 ± 0.36 2.57 ± 0.37 2.41 ± 0.25Concrete 16.07 ± 5.75 8.96 ± 1.73 3.22 ± 0.31 3.21 ± 0.26 3.06 ± 0.18Energy 9.54 ± 4.54 6.70 ± 2.39 1.61 ± 0.40 1.51 ± 0.28 1.38 ± 0.22Kin8nm 2.12 ± 0.97 0.11 ± 0.20 -1.11 ± 0.04 -1.12 ± 0.04 -1.20 ± 0.02Naval propulsion plant -5.68 ± 0.34 -5.85 ± 0.15 -5.65 ± 0.28 -4.08 ± 0.13 -5.63 ± 0.26Power plant 35.78 ± 12.87 22.04 ± 4.42 2.82 ± 0.04 2.82 ± 0.04 2.79 ± 0.04Protein 40.98 ± 7.43 25.73 ± 1.59 2.87 ± 0.03 2.91 ± 0.03 2.83 ± 0.02Wine 33.73 ± 10.75 20.55 ± 3.72 1.95 ± 4.08 1.58 ± 2.30 0.94 ± 0.12Yacht 10.18 ± 4.86 6.85 ± 2.84 1.26 ± 0.29 1.28 ± 0.36 1.18 ± 0.21Year Prediction MSD 39.02 ± nan 21.45 ± nan 3.41 ± nan 3.39 ± nan 3.35 ± nan

Table 2: Additional results on regression benchmark datasets: the top table reports RMSE and bottomtable reports NLL. Ensemble-M (MSE) denotes ensemble of M networks trained to minimize meansquared error (MSE); the predicted variance is the empirical variance of the individual networks’predictions. ML-1 denotes maximum likelihood with a single network trained to predict the meanand variance as described in Section 2.2.1 . ML-1 is significantly better than Ensemble-5 (MSE) aswell as Ensemble-10 (MSE), clearly demonstrating the effect of learning variance. ML-1+AT denotesadditional effect of adversarial training (AT); AT does not significantly help on these benchmarks.ML-5, referred to as deep ensembles in Table 1, is an ensemble of 5 networks trained to predict meanand variance. ML-5+AT results are very similar to ML-5 (the error bars overlap), hence we do notreport them here.

A.2 Effect of training using MSE vs training using NLL

A commonly used heuristic in practice, is to train an ensemble of NNs to minimize MSE and use theempirical variance of the networks’ predictions as an approximate measure of uncertainty. However,this generally does not lead to well-calibrated predictive probabilities as MSE is not a scoring rulethat captures predictive uncertainty. As a motivating example, we report calibration curves (alsoknown as reliability diagrams) on the Year Prediction MSD dataset in Figure 7. First, we computethe z% (e.g. 90%) prediction interval for each test data point based on Gaussian quantiles usingpredictive mean and variance. Next, we measure what fraction of test observations fall within thisprediction interval. For a well-calibrated regressor, the observed fraction should be close to z%. Wecompute observed fraction for z = 10% to z = 90% in increments of 10. A well-calibrated regressorshould lie very close to the diagonal; on the left subplot we observe that the proposed method, whichlearns the predictive variance, leads to a well-calibrated regressor. However, on the right subplot, weobserve that the empirical variance obtained from NNs which do not learn the predictive variance(specifically, five NNs trained to minimize MSE) consistently underestimates the true uncertainty.For instance, the 80% prediction interval contains only 20% of the test observations, which meansthe empirical variance significantly underestimates the true predictive uncertainty.

13

0.0 0.2 0.4 0.6 0.8 1.00.0

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.00.0

0.2

0.4

0.6

0.8

1.0

Figure 7: Calibration results on the Year Prediction MSD dataset: x-axis denotes the expected fractionand y-axis denotes the observed fraction; ideal output is the dashed blue line. Predicted variance (left)is significantly better calibrated than the empirical variance (right). See main text for further details.

B Additional results on classification

B.1 Additional results on MNIST

Figures 8 and 9 report results on MNIST dataset using different architecture than those in Figure 2(a).We observe qualitatively similar results. Ensembles outperform MC-dropout and adversarial trainingimproves performance.

0 5 10 15Number of nets

1.0

1.1

1.2

1.3

1.4

1.5

1.6

1.7

1.8Classification Error

Ensemble

Ensemble + R

Ensemble + AT

MC dropout

0 5 10 15Number of nets

0.02

0.04

0.06

0.08

0.10

0.12

0.14NLL

Ensemble

Ensemble + R

Ensemble + AT

MC dropout

0 5 10 15Number of nets

0.0015

0.0020

0.0025

0.0030

Brier Score

Ensemble

Ensemble + R

Ensemble + AT

MC dropout

Figure 8: Results on MNIST dataset using the same setup as that in Figure 2(a) except that we use twohidden layers in the MLP instead of three. Ensembles and adversarial training improve performanceand our method outperforms MC-dropout.

0 5 10Number of nets

1.3

1.4

1.5

1.6

1.7

1.8

1.9

2.0

2.1Classification Error

Ensemble

Ensemble + AT

MC dropout

0 5 10Number of nets

0.04

0.05

0.06

0.07

0.08

0.09

0.10

0.11

0.12

0.13NLL

Ensemble

Ensemble + AT

MC dropout

0 5 10Number of nets

0.0020

0.0022

0.0024

0.0026

0.0028

0.0030

0.0032

0.0034

Brier Score

Ensemble

Ensemble + AT

MC dropout

Figure 9: Results on MNIST dataset using a convolutional network as opposed to the 3-layer MLP inFigure 2(a). Even on a different architecture, ensembles and adversarial training improve performanceand our method outperforms MC-dropout.

B.2 Qualitative evaluation of uncertainty

We qualitatively evaluate the performance of uncertainty in terms of two measures, namely theconfidence of the predicted label (i.e. maximum of the softmax output), and the disagreement betweenthe ensemble members in terms of Jensen-Shannon divergence. The results on known classes are

14

shown in Figures 10 and 11. Similarly, the results on unknown classes are shown in Figures 12(a)and 12(b). We observe that both measures of uncertainty capture meaningful ambiguity.

Figure 10: Results on MNIST showing test examples with high or low disagreement between thenetworks in the ensemble: Top two rows denote the test examples with least disagreement and thebottom two rows denote test examples with the most disagreement.

Figure 11: Results on MNIST showing test examples with high or low confidence: Top two rowsdenote the test examples with highest confidence and the bottom two rows denote test examples withthe least confidence.

7 1 3 7 3 1 2 1 7 3

3 1 3 1 1 3 1 3 1 1

7 9 0 9 0 8 0 9 8 0

0 4 4 5 0 8 0 8 4 3

(a) Least and most disagreement

3 3 3 3 3 3 0 3 3 2

3 3 3 3 3 3 2 3 3 3

5 0 7 3 7 0 6 2 5 4

6 5 7 6 4 0 5 0 5 8

(b) Highest and least confidence

Figure 12: Deep ensemble trained on MNIST and tested on the NotMNIST dataset containing unseenclasses: On the left, top two rows denote the test examples with highest confidence and the bottomtwo rows denote the test examples with the least confidence. On the right, top two rows denote thetest examples with highest confidence and the bottom two rows denote the test examples with theleast confidence. We observe that these measures capture meaningful ambiguity: for example, theensemble agreement is high for letters ‘I’ and some variants of ‘J’ which resemble 1 in the MNISTtraining dataset. The ensemble disagreement is high and confidence is low for examples visuallydis-similar from the MNIST training dataset.

15

1 Introduction2 Deep Ensembles: A Simple Recipe For Predictive Uncertainty Estimation 2.1 Problem setup and High-level summary2.2 Proper scoring rules2.2.1 Training criterion for regression

2.3 Adversarial training to smooth predictive distributions2.4 Ensembles: training and prediction

3 Experimental results3.1 Evaluation metrics and experimental setup3.2 Regression on toy datasets3.3 Regression on real world datasets3.4 Classification on MNIST, SVHN and ImageNet3.5 Uncertainty evaluation: test examples from known vs unknown classes 3.6 Accuracy as a function of confidence

4 DiscussionA Additional results on regressionA.1 Additional results on regression benchmarksA.2 Effect of training using MSE vs training using NLL

B Additional results on classificationB.1 Additional results on MNISTB.2 Qualitative evaluation of uncertainty